You Only Look Once Based-C2fGhost Using Efficient SIoU Loss Function for Airplane Detection

Detecting airplanes in satellite imagery presents significant challenges due to the intricate backgrounds and varying conditions of data acquisition influenced by sensor geometry and atmospheric effects. Although deep learning algorithms are rapidly advancing, their primary application and evaluation have been on widely-used ground-based imagery. This research provides a comprehensive evaluation and comparison of several sophisticated object detection algorithms tailored for aircraft identification in satellite imagery. By utilizing the extensive HRPlanesV2 dataset alongside a stringent validation process on the GDIT dataset, we train a state-of-the-art object detection model using YOLOv10. Furthermore, this research proposes YOLOv10-C2fGhost, which improves the C2f module to C2f-Ghost, further reducing model complexity. Additionally, the replacement of the SIoU loss function with CIoU in YOLOv10s demonstrates significant improvements in accuracy and detection of small objects. On the GDIT dataset, experimental results show that, compared to the original YOLOv10s, the number of parameters was reduced from 8M to 6.6M, and GFLOPs decreased from 24.4 to 17.2, while maintaining a similar accuracy of 92.2%. Similarly, experiments on the HRPlanev2 dataset showed a significant reduction in model complexity and a high accuracy of over 97%.